‍

A growing team of 40 with experts in every function—including finance, legal, operations, comms, platform and our proprietary Riskgaming—support an expanding investment team of technical generalists with divergent passions and investing styles converging under the banner of “Unum Lux” (one Lux) to compete and earn the right to win with founders, believing before others understand.

Amid a busy few months, we have pursued some of the highest-quality founders and opportunities in our history, investing more than $450 million across over 70 new and existing companies with another $100 million across 12 more companies closing in the coming weeks. Lux’s opportunity set has long focused not on singular breakthroughs but on the convergence of computational and concrete capabilities across previously disconnected domains to transform atoms as fluidly as bits.

We unearth many of these opportunities by charting the directional arrows of progress across a variety of sectors—cheaper computation; higher energy density per unit of raw material; human-computer interfaces exchanging tactile keyboards for gestures captured by reading neuromuscular signals; declining payload costs to access space; rapidly improving and asymmetrically inexpensive autonomous and attritable defense systems—and then heading to where those arrows converge.

We caught some of these converging arrows by founding or being the first or early investor in exited companies like Kurion, CTRL-Labs, Auris Health, MosaicML, Recursion, Matterport, Aeva and Evolv. In addition, we’ve invested in new convergences in high-demand, high-confidence winners including Anduril, Hugging Face, Applied Intuition, Saildrone, Runway, Eikon, Lumafield, Varda, Hadrian, Databricks, Ramp, Together AI, eGenesis, Impulse Space, Nominal, Osmo, Sakana, EvolutionaryScale, Cambridge Aerospace, Covenant Industries, Variant Bio, Physical Intelligence, Kela, Amca, and Cognition. These realized and unrealized gains total billions of dollars for our investors.

Lux works to earn the leverage of long. Traders obsess over the tick, journalists over the day and analysts over the quarter. The essence of our "Five-year Psychological Bias" is that everyone wants to be invested today where they should have been looking five years ago. By orienting capital toward 2030‑scale inflections, we continually harvest an edge others might not or won’t bankroll. Conviction compounds by trading the panic of the moment for patience of the decade, increasingly fleeting against a barbell backdrop of small firms (“the minnows”) involuntarily exiting via shutdowns and large firms (“the megas”) voluntarily planning to exit via IPO.

For investors, the edge remains temporal: when the crowd trades the minute, bet the decade; when attention crowds the valley, climb the ridge. Though as we’ll discuss shortly, temporal edge eventually meets physical edge: every big idea eventually rams into biology, energy or thermodynamics—what we can call the Friction Frontier.

Attention is all you need in this AI frenzy—if you can secure it
Einstein once said that the only reason for time is so everything doesn’t happen at once. Attention is our most precious resource; it’s also attenuating as virality virtuosos ping pockets with push alerts to be loved or loathed—but never ignored. A high-velocity slipstream of science, markets and geopolitics now batters even the most stoic students of history. Time today, to Einstein's presumed chagrin, can feel like a traffic jam of simultaneous emergencies.

Attention is also the most important resource for artificial intelligence. Coincidentally, it was a founder of Lux-family company Sakana AI who titled the famous "Attention Is All You Need" paper while still at Google, co-authoring the invention of transformers which transformed AI into today’s all-consuming phenomenon.

AI is counter-proof to Einstein: time doesn't exist, since everything seems to be happening at once. Megas and minnows, craftsmen and charlatans, technicians and tourists alike are all accelerating their activity—even as they abjectly abbreviate their attention. We will first focus on the heady headlines before swiveling and zooming our lens toward the quieter periphery where fresh opportunity awaits for those with practiced patience.

At the infrastructure layer, the capex surge underwriting AI models is already soaking up roughly 5% of U.S. GDP, placing it shoulder-to-shoulder with the fiber-optic and router binge of the 2000 tech boom (5.2%), and only a short step below the sheet-rock bonanza of the 2005 housing bubble (6.7%). In its favor, capital today isn’t being entombed in passive drywall; it’s being wired into actively self‑improving, power‑hungry nervous systems that learn, iterate and compound—even as they depreciate. Yet, history suggests euphoric investors will overspend on AI capex––and it remains to be seen who captures the derivative value (more prudent investors, companies or customers, though increasingly likely in that order).

At the model layer, the AI market has split into two: closed models (OpenAI, Anthropic and Google) charging for access and already generating revenue in the billions of dollars versus open-source alternatives (Meta's Llama, Mistral) that now perform nearly as well as their pricier cousins. This changes the investment calculus. The closed players enjoy 50-70% gross margins, but rapidly burn their revenues and new capital raises on training their models to capture every edge possible. Meanwhile, open-source platforms can be more limited in their revenue potential, but scale efficiently with good-enough performance and unbeatable cost-per-token inference. Smart money may own both sides—equity in the walled cathedrals and positions in the open bazaar.

That refined thinking has led to frantic strategic repositioning as companies race to offer differentiated options. OpenAI, which hasn’t released an open-source or open-weights model since GPT-2 in 2019, just launched gpt-oss (hosted on Lux-family company Hugging Face), an open-weights model it claims achieves parity with its o4-mini proprietary model. Coming the other direction, Meta has starkly changed its course, rebuilding its superintelligence team through what we’ll call “Poachapalooza 2025.” Today, Meta can offer an individual AI scientist $1 billion to advance their capabilities—and that individual will turn it down to wait for a better opportunity.

Due to the FTC and the DOJ’s determined deterrence to deal-doing, we are witnessing the rise of “L&A”—license and acqui-hire—first pioneered by Microsoft with its acquisition of Inflection, followed by Amazon with Adept and Alphabet with Character.ai, and now perfected by Meta with Scale AI. It’s dirty dancing around the intent of Delaware corporate governance and a serious threat to America’s innovation economy. We expect future term sheets will contain clauses to protect stakeholders from 'talent raids,' 'partial acquisitions,' and 'orphaned equity,' all vigilantly erected to guard the near future against present exploits endured vulnerably in the recent past.

There’s a great irony in the current frenzy for artificial intelligence: What is being acquired is not hardware or contracts, but human intelligence. The twist is that AI hands these humans enormous leverage: two engineers with Gemini 2.5 Pro, Claude 4 Opus, Grok 4 or ChatGPT o3-pro may now outmaneuver a hundred-lawyer firm or a legacy consultancy, because the playing field now depends more on implementation skill than headcount. Every white-collar profession faces the same reckoning: yesterday's moats may be today's liabilities, and tomorrow belongs to whoever best orchestrates the machines. Are firms placing enough attention on their own impending demise? Time may have stopped against Einstein’s view, but it continues to flow mercilessly forward nevertheless.

Attending to the verified truth of AI’s agentic future
Pop the hood on today’s intelligence engine and you’ll find the choke point isn’t silicon horsepower—those metronomic doublings are now table stakes—but instead labeled nuance, the rare isotope of modern cognition.

The frontier of agentic AI capability today isn't defined by model size—but between what we can verify and what we can't. We can train an AI model to solve differential equations because every answer resolves to binary truth: correct or incorrect. But ask that same AI to write a compelling opening to a novel, and it gets lost in a wilderness of taste—is "It was a dark and stormy night" genius or cliché? Is “Call me Ishmael” immortal or just terse? AI, like Walt Whitman, contains multitudes. It can be superhuman at chess and protein folding yet easily fumble at tasks of taste. It isn’t that verification is hard—it's that verification is today’s rate-limiting step for AI progress.

Matters of taste are much more expansive than just generating cultural artifacts. Despite 5% of American GDP devoted to AI data infrastructure, our AI agents still can't book a flight from JFK to SFO with precision. Think of the preferences involved: which time of day, aisle or window, how much to gamble on delays, cash or frequent flyer points, the likelihood of that elusive first-class upgrade. Those preferences aren’t isolated either: would we prefer a later flight if we could have an aisle seat and a better chance of an upgrade? The math is brutal—even a 99% reliable agent degrades to 60.5% accuracy after just 50 sequential decision tasks, each error cascading through the system like spidering cracks in tempered glass.

In domains where we can construct tight feedback loops—code that compiles, games with win conditions, molecules that bind—AI can move towards mastery. But in the vast expanse of human endeavor where "good" is contextual or cultural, there is still space for humans.

Companies are voraciously solving for verification. Meta’s raid on Scale AI was a smash‑and‑grab for labeled nuance, securing the data infrastructure necessary for its next-generation, proprietary models. OpenAI, Anthropic and even xAI’s Grok are all quietly erecting their own annotation foundries, atomizing every step in math and physics Olympiad proofs because tomorrow’s edge will be measured in insight per token, not teraflops per second. Multi‑agent reinforcement learning verifiers now grade everything from courtroom prose to manufacturing workflows, turning the subjective into back‑propagatable currency.

This evolution from static data to agentic learning environments marks a critical shift in the directional arrow of future progress in AI. Together AI’s record‑setting inference engine on Nvidia’s Blackwell architecture is unmatched, offering engineers the fastest loop for learning and accelerating agent training. That speaks to Lux-family company Cognition and its Devin coding agent, which is now the leading copilot for the best software engineers in the world. Devin’s speed of development is only matched by Cognition’s own rapid maturation. Their reconnaissance and rapid fielding of the Windsurf team was a masterful move making us even more bullish on their smart, fast-thinking and long-term trajectory.

Sakana AI’s rollout across three Japanese megabanks shows the leap from toy problems to high-finance value as its models rapidly evolve—just as Sakana uses nature’s evolutionary patterns to inspire its own research. Ramp solved agentic error with surgical scaffolding—when their AI features misstep, the core infrastructure stays bulletproof. Tldraw abandons the typical reliance on the chatbot paradigm entirely, reimagining AI through brush strokes rather than prompts—showing that sometimes the user interface itself is the best site for innovation. Finally, sometimes different types of labeled nuance come together to create advantage. Multimodality is crushing the latency between imagination and incarnation. Runway’s sold‑out AI Film Festival at Lincoln Center—exhibiting films dreamed up on its own tech stack for generative video—proves that a prompt and its prompter can now walk the red carpet.

As these examples show, instead of just spending more on compute in the pursuit of ever-bigger models, there is arbitrage in transmuting the subjective into the scoreable to help models attend to verified truth. The winners might not be those with the most parameters, but those who architect new verification primitives and make legible areas currently invisible to AI.

The Friction Frontier: Investing in hard assets
Friction is the first villain of physics, the thief of efficiency that steals energy from every moving part, making our engines run hot and our transmissions grind. We spend fortunes on ball bearings and lubricants, on streamlining and smoothing, on fiber optic cables that carry light itself because electrons aren’t fast enough. In war, friction kills—every second of delay between order and execution measured in lives.

The wealthy have always understood that money's greatest luxury isn't what it can buy but what it can bypass—the private jet that leapfrogs TSA lines; the concierge doctor who answers on the first ring; the Disney plaid-vested guide who spirits you past the sweating masses. It's not mere snobbery that drives this parallel infrastructure of privilege; it's the recognition that friction is modernity's hidden tax, levied in time and aggravation. That’s why the ultimate arbitrage lies in friction—owning the bottlenecks imposed by physics and funding the hacks that beat them.

There’s a paradox though: the most elegant physical systems often depend on friction as their saving grace. Nuclear reactors use control rods to slow neutron flow—too little friction and Chernobyl burns; the right amount and cities glow with light. Our neurons fire with built-in delays, synaptic gaps that prevent our thoughts from becoming electrical storms of epileptic seizure. In our very cells, DNA's checkpoint proteins pause replication to proofread and mend molecular misprints, because without that brake, every miswrite might script a tumor. The question isn't whether we need friction, but which kind, where and how much we can bear before we pay someone to help us escape it—what we call the Friction Frontier. If friction is sometimes a virtue, then we must be surgical about where we apply the brake.

Our investment instinct is thus two‑fold: own both the bottleneck and the hack by backing the jailbreakers. Take artificial intelligence. The Friction Frontier lies in mapping AI's metabolism, finding where silicon speed meets carbon constraint. AI can dream up chip architectures in microseconds but TSMC's fabs run on geological time, their extreme ultraviolet lithography crawling through wafers at the pace of physics. AI can generate perfect building blueprints, but concrete still cures at its own cadence. The rate-limited steps are hiding in plain sight: rare-earth mining that feeds the hunger for GPUs, the electric grid that can't be coded into existence, the submarine cables that must be unspooled from a vessel one meter at a time.

One can be long Caterpillar's earthmovers, ASML's ultraviolet monopoly or whatever can't be software-defined in the cloud. But it’s also important to be long the lubricants—the companies building the tools that let AI route around reality's roadblocks, the robotics ventures that give AI hands to match its mind. Future fortunes lie at the Friction Frontier: profiting from what can't be rushed and also what makes rushing possible.

In robotics, we see a careful convergence of the need for validated data while staying within the limits of physics. There's a tension between what we want (powerful robots that can handle real-world complexity) and what's practical to build (collecting robot data is expensive and slow). We wanted to own the bottleneck on real-world data while backing the jailbreaks behind a hack, so we bet on Physical Intelligence (π). The company is building a foundation model and operating system platform for improvisational robots that can enter a highly unknown and unstructured environment and perform tasks loosely defined. It’s built a proprietary workflow to collect the largest robotics dataset in the world, and it’s now building flexible physical world models that will challenge our perception of robots unable to handle the most mundane of tasks. Laundry will never be the same again.

That same imperative to embody intelligence in the physical world leads us from silicon to the Friction Frontier of the cell. Indeed, AI platforms for protein, small molecule and drug design are emerging from the thawing ice of the biotech winter, accelerating new drug candidates and eliminating the inefficiencies of clinical trials. All are gaining interest, attention and capital but not yet full-fledged FOMO—another anticipated example of our Five-year Psychological Bias.

The real revolution hides in plain sight: biotech is pivoting from fixing broken people to upgrading functional ones. Variant finds outlier people with outlier traits in outlier places of the world to target not just medicines but also performance enhancers for anyone. Its addressable market just exploded from "patients" to "everyone who wants to be better." The industry is shifting from treating the sick to optimizing the well, from one-size-fits-all drugs to N-of-1 personalized therapeutics, from managing decline to engineering longevity. Biotech’s poor public market performance (-23% over five years versus the S&P's +95%) masks a private market truth: we're seeing a Darwinian moment where yesterday's risks may become tomorrow's moats because patience, correctly applied with enough funding, turns time into an asymmetrical advantage.

We observe this same convergence on the Friction Frontier across Lux’s companies. Eikon’s industrial‑scale microscopes fuse picometer precision with high‑throughput automation and AI‑rich analytics—watching proteins dance in living cells and shrinking years of drug‑discovery guesswork into months. Impulse Space crafts orbital‑transfer vehicles that accelerate Mars timelines from decades to months, while Applied Intuition spins digital‑twin asphalt grids that turn millions of on‑road driving miles into an overnight compute sprint. Hadrian’s robotic foundries attack the meta‑manufacturing choke point—building the machines that build the machines, while Varda’s hypersonic re‑entry labs collapse orbital experiment cycles from months to days. Anduril is developing autonomous fighter jets at record speed, while EvolutionaryScale compresses 500 million years of evolution into hours, conjuring a green fluorescent protein that nature hadn’t yet found a path to. Friction may be the first villain of physics, but it also offers us a redemption story—if we stay focused and listen closely.

Crowd consensus and our diverging views
To deviate from consensus, we must know the consensus and how everyone else directs their attention. Below are our views on the daily distraction tape—the headlines better to arb than echo.

As of writing, the S&P keeps minting fresh peaks while investment‑grade spreads shrink to a mere ~80 bps over Treasuries and investors pay over 3x sales for the median name—an all‑time high. Retail investors poured a record $155 billion into U.S. equities in the first half of this year while over 850 unprofitable Russell 3000 companies averaged 36% gains since April. Nine Big Tech stocks make up 36% of the entire S&P index—a half-century high—concentrating risk.

There’s increasingly crowded consensus on public AI names alongside the return of ‘Hot Meme Summer’ with GoPro, Krispy Kreme and other darlings of retail speculators in equity-euphoria, while Bitcoin hit recent record highs of more than $120,000. Alas, physics teaches us that excess energy must eventually dissipate and so too must frothy finance. As capital’s cost reasserts its gravity, we can foresee a scenario where narratives may shift from growth glitz to maintenance mettle. Today’s record AI capex expenditures on chips, data centers, nuclear reactors and turbines will succumb to wear-and-tear due to physical entropy, creating future demand for predictive sensors, software and robotics for prevention and repair.

In private markets, the venture asset class doubled its market share to 22% in the secondary liquidity market. An interesting, non-obvious signal that possibly portends an opening of the IPO market for high-demand private companies is that the asset class had the sharpest pricing gains, reaching 78% of NAV. Meanwhile, retail-sourced secondaries surpassed $80 billion (doubling in the past two years), adding more heft to this new liquidity infrastructure. It all means sophisticated (and less so) secondary buyers can be patient long-term holders who now underwrite late-stage venture based on company fundamentals rather than exit timing or distressed sellers. We continue to believe that regardless of the IPO market, investing in must-have teams and technologies in markets with oligopolistic industry structure can foment bidding competition that leads to rationally irrational prices to buy our and our founders’ stakes.

Internationally, Israel has destroyed or degraded the majority of Iran’s proxies existentially threatening its sovereignty: Houthis, Hezbollah, Hamas and an extraordinarily precise decimation––in a mere 12 days––of Iranian air defenses with a multi-year setback of its offensive nuclear intentions thanks to a key assist from CENTCOM and the technological superiority of American B-2 stealth bombers. Israel’s young warfighter generation will likely be its greatest generation. Lux partners have traveled to Tel Aviv and we have made four core investments in extraordinarily intelligent and talented teams in the past year alone spanning defense and cyber relevant not just to Israel but also the United States and the free and democratic world more broadly.

On that democratic world, the ideological pendulum swung between extremes: populist right and socialist left evident in France, the United Kingdom and Germany. In America, we’ve seen military deployments in cities like Los Angeles, immigration raids as well as universities facing massive funding cuts for failing to protect the safety of their students while ensuring both free speech and orderly dissent. Lux has long partnered with the best principal investigators and academic scientific entrepreneurs facing funding cuts and loss of critical talent to offer resources and help continue key experiments. We launched the Lux Science Helpline to step in and step up, fielding proposals from some of America’s next generation of leading scientists. We offered them advice, connections and—if a fit for our funds—funding to launch new companies. Where appropriate, we also facilitated introductions to existing Lux family companies who could recruit them or license their work.

Culturally, Coldplay’s high-fidelity stadium kiss‑cam revealed on-screen infidelity, becoming a high-circus slapstick parable of always‑on surveillance, a reminder that terrible judgment can be alchemized into terribly addictive meme gold and that human creativity can spin personal tragedy into communal comedy. Yet, we can see a near-future where “lifecording” becomes ambient and assumed. Consider long-battery wearables that passively capture all audio, glasses that passively record and capture what you “see” even when you aren’t paying attention, and eventually via Osmo, sensors capturing the smells, chemicals and molecules you encounter every day. Now add in a layer of AI revealing insights and hidden connections to make sense of things that evade your attention or recall, and this “lifecording” phenomena is not far off. Like earlier debates over social media, the younger, digitally native generation will be far more welcoming and comfortable with this new world while elders fear a further invasion of their already-diminished privacy.

Markets, politics and culture toggle between rewind and fast‑forward: yesterday’s messages re‑skinned for today’s nostalgia, today’s absurdities served for tomorrow’s algorithmic clicks. The only forecast that feels safe: more remix, less middle ground and a perpetual beta test of our collective focus. What's always missing—and thanks to collective human ambition, will always be absent––is a "pause" button.

The Friction Frontier in the Age of Superintelligence
With superhuman intelligence imminent, where does this leave all of us humans? The irony is that the cultural critic can find comfort in their continuity; the bedside nurse with grace and humanity accentuates the second syllable of healthcare; the blue-collar worker who we all thought would be most disrupted by the march of technology can firmly plant their feet with their finely honed skills and judgment. Meanwhile, the white-collar worker is sweating because they know they are vulnerable—and doubly vulnerable if they’re not sweating.

The shrewdest strategy is going long on both the friction and the lubricant—owning what won't budge and also what might break it loose. It's investing behind both mortality and medicine. We accept that bodies decay on schedule, and so we invest in healthcare IT that coordinates care. But we’re simultaneously long CRISPR's cut-and-paste potential, Aera's genetic cargo delivery that slips medicine past cellular customs, and eGenesis's porcine spare parts grown in spotless labs. We don't fight the fact that organs fail, but we don’t dismiss the bio-hackers engineering kidneys and pig hearts that may soon beat in human chests. We fund their blasphemy against biology's deadlines.

Or consider being long both darkness and daylight—accepting that circadian rhythms rule our commerce and our cortisol. Yet we’re also long Reflect Orbital's audacious mirrors, space-based spotlights that would bounce sunbeams past the planet's bedtime, selling sunshine after sunset to cities and solar farms willing to pay for one more hour of productivity and one more degree of warmth and energy. The bet isn't on conquering Earth’s rotation—it's on profiting from both its spin and the schemes to transcend it.

In this upcoming age of AI superintelligence, what becomes of venture capitalists themselves? Here, the ghost of Arthur Schopenhauer has hovered over Lux partner meetings for years and continues to inform us. “Talent hits a target no one else can hit; genius hits a target no one else can see.” Lux has made its rep and raison d'etre on that invisible arc: finding and funding founders with arrogance of the highest order, believing before others understand, staking the outcast idea far before the analyst memo. Belief precedes understanding; capital chases conviction. Where others might need proof, Shopenhauer's ghost whispers possibility—so we focus where he does, on those founders inventing futures still invisible to the rote pattern-recognition crowd.

Even in the age of superintelligence, the deepest fortunes aren't made by picking sides in the ancient wars of physics but by funding both the immovable object and the unstoppable force. Turning the impossible into the inevitable. In a universe governed by entropy and ingenuity in equal measure, winning portfolios profit not from friction's elimination but from humanity's eternal negotiation with the Friction Frontier. In a world where everyone shouts at once, we invest in the few special voices that still whisper decades ahead. Fiat Lux.